Since I don’t have enough dumb war training data, I’m not trusting my intraday training models. So it’s switched off until these thresholds resume to normal.

Interestingly, the last time there was this much uncertainty was in the knucklehead tariff era, where so-called Liberation Day was quickly followed by Capitulation Day, when Trump TACOed out due to the inevitable consequences of this ill-considered policy, which was rollbacked a couple of months later (see yellow band in middle of this 2 year view)

Hasta luego para ahora

A

This administration does not lend itself to geopolitical strategy. Analysis from commentators and professors is that there simply hasn’t been an president or cabinet this dumb to fall into a trap like this for a long time.

It makes me nostalgic for the 90s thrillers where dark, shady, Machiavellian schemers preserved US supremacy. Now we can’t watch those those movies with credulity. We can’t watch Clear and Present Danger without a wry smile…. our present generation of wrongdoers are just not that smart.

Beyond the countless illegal endeavors of this administration, there are also a few of questionable morality.

Unusual Machines

This company caught my eye when Money Master made a recent purchase.

Looking historically, I saw a +400% gain a few weeks after their Q3 2024 call. Before this point Unusual Machines was a penny stock searching for an audience. The press releases tell the story of a company scraping for attention.

This was a company whose most significant milestone was getting a flight controller approved for the Blue UAS Framework, a U.S. Department of Defense program for vetted drone components - an impressive achievement, which barely moved the price until sharp drop two months later.

However, good fortune was in store towards the end of 2024.

“Shares of drone maker Unusual Machines soar after Donald Trump Jr. joins advisory board” was the reason. The smart money got in on the name, and the dumb money got in on the articles about the name.

Regardless, the stock has continued to do well overall to the present day.

Unusual Machines is now acquiring companies, signing defense supplier agreements, and leasing a 17,000 sq ft motor factory in Orlando.

And naturally has been streamlined onto a Pentagon contract in October 2025 for drone parts. 

Eric Trump and Donald Trump Jr. also backed a merger that folded their golf-linked Aureus vehicle into Powerus, a military-drone company pursuing Pentagon contracts that tap into a new policy that bars the Defence Department from procuring or operating drones that are made in a covered foreign country, e.g. China.

Amazing timing for them to diversify from golf into defense hardware.

America is a socialist country at the top and a free market at the bottom.

Be well,

Alex

But… this is just the beginning

In the last post I got to a decent 0.60 AUC baseline with binary classification machine learning i.e. can the system accurately guess if the stock will go up or down.

However, I had some operational issues, and probably the biggest thing is we have a full-blown war going on right now for which I just don’t have training data since we haven’t had a war as big or dumb as this for a long time.

However… with the back test I wanted to see if I could push model performance and with the help of Opus 4.6 tested some new ideas.

Now for some quick experiments to increase the AUC and win rate.

Experiment 1: Remove ‘post event’ metrics

The audit flagged that delta_premarket, delta_transcript, delta_gap, premarket_volume_pct and related features happen AFTER the earnings call. They’re valid if if we trade at market open when the premarket reaction is known, but they create two problems:

• They dominate the model — delta_premarket_positive is currently the #1 feature in the 4h model

• They conflate “the model learned something from the transcript” with “the market already reacted”

We trained two XGBoost EOD models to compare the effect of premarket/gap features on prediction accuracy:

1. With premarket features (the standard model) — uses 192 features including gap_pct, delta_premarket, premarket_volume, etc. Result: Test AUC 0.627, 70.3% win rate at 0.70 threshold, +1.95% net avg return.

2. Without premarket features — drops 27 post-event price features so the model only uses transcript/LLM signals, fundamentals, and historical data. Result: Test AUC 0.552, 61.0% win rate at 0.70 threshold, +0.82% net avg return.

The point here is that a no-premarket model could let us place trades before the market opens (out-of-hours), as soon as the earnings transcript is available — we wouldn’t need to wait for premarket price action. The tradeoff is a significantly thinner edge (~7.5 AUC points lower) though.

Prices are less liquid out of hours so this is risky territory but perhaps worth pursuing.

Both are set up as analysts that do not trade so we can paper test.

Experiment 2: Feature Pruning & Premarket Experiment

A new model was trained by pruning from 192 features down to just 23.

This aggressive pruning improved test AUC from 0.627 to 0.633 and win rate at the 0.70 confidence threshold from 68.6% to 71.6% (+2.28% net average return per trade, 225 trades).

The simpler model is less prone to overfitting noise and performs better across all market cap segments, with large-cap AUC at 0.656 and mid-cap at 0.636.

The most powerful signal by far is the premarket price reaction — delta_premarket_positive alone has nearly 2.5x the importance of any other feature. Combined with the raw premarket delta, transcript-period price move, and the 50-day EMA relative to premarket price, market reaction features dominate the model. This confirms that the market’s immediate overnight response to an earnings call is the strongest predictor of where the stock closes that day.

Beyond price action, the model relies on a company’s earnings track record (win rate, consecutive beats, typical move size), LLM-extracted transcript signals (management confidence, new negatives, backlog weakness, cost control, guidance beats), timing (before vs after market), and a handful of industry-specific effects (Semiconductors, Restaurants, Biotech, Solar behave differently post-earnings).

Market context like VIX and market cap category are also important.

The key insight is that 23 features carry virtually all the predictive signal — the other 169 were adding noise, not information.

Experiment 3: Temporal feature weighting

Give more weight to recent quarters vs older ones. Market regimes change — a model that overweights 18-month-old patterns may underperform on current conditions.

We added temporal recency weighting to the training pipeline that applies exponential decay to sample weights based on how old each training quarter is. With a 9-month half-life, the most recent quarters get full weight while the oldest training data (from early 2024) is down-weighted to about 38% influence. The idea is that more recent earnings patterns are more predictive than patterns from 18+ months ago. We tested half-lives of 6, 9, and 12 months against a no-recency baseline on XGBoost EOD, and 9 months was the clear sweet spot — improving test AUC by +0.012 over baseline at the same iteration count.

We then combined recency weighting with the feature pruning approach to produce new v8 models across all four targets.

The results: XGBoost EOD landed at 0.631 test AUC with 72.7% win rate at the 0.70 threshold (+2.48% net after 35bps costs), LightGBM EOD hit 0.631 with an impressive 76.0% win rate at 0.65, XGBoost 7d improved significantly to 0.624 (up from 0.608 in v6), and LightGBM 28d reached 0.632 with 70.9% win rate and +8.03% average net return per trade.

Experiment 4: Walk-forward validation instead of single train/test split

Right now I have one split (e.g. train before Oct 2025, test Oct 2025 – Jan 2026). This is a single window that could be unusually easy or hard.

Implementing rolling walk-forward (e.g., train on 12 months, test on the next 2 months, slide forward) gives 5-6 evaluation windows and much more robust AUC estimates.

The findings were:

1. XGBoost 4h and XGBoost EOD are the strongest — both have pooled AUC around 0.605-0.608, confirming genuine short-horizon signal

2. Signal degrades with longer horizons — 7d drops to 0.579, 28d to 0.557. This makes intuitive sense; earnings-driven alpha fades as macro/sector noise dominates

3. All models show the same pattern — weak early folds (0.55) when training data is sparse, strong later folds (0.63-0.66) once 10k+ samples available. More data = better signal, might need to go back further in time.

4. CV AUC is consistently ~0.62 across all models, while test AUC varies more, confirming some regime sensitivity

5. Despite modest AUCs, the trading economics work — 60-67% win rates at high thresholds with positive net returns after costs

6. XGBoost consistently edges out LightGBM on walk-forward test AUC across all horizons

The walk-forward results validate that the currently trading XGBoost EOD model is indeed best.

Experiment 5: Different time horizons

Currently I try and predict EOD prices, but there are arguments for…

Closing sooner e.g. 2h, 3h, 4h. Most of the bump in good earnings happens at the start of the day, so close out sooner.

Or closing later - in the last post I mentioned that 28d models performed best, if I can find a capital efficient way of doing this it might be a better route.

• 3d is the best AUC at 0.636 — a new horizon we hadn’t trained with v8 before.

• 28d has the highest returns by far — +7.64% net avg at 0.70 threshold with 70.4% win rate (351 trades). The longer holding period allows the edge to compound.

• 2h and 4h are very similar in AUC (~0.62) but the high-confidence thresholds (0.75-0.80) produce excellent win rates of 72-80%.

Experiment 6: Add options-derived features

Added options features below, sadly it’s not possible to backfill these from my current broker so we will have to experiment going forward.

Experiment 7: Predict magnitude, not just direction

Instead of binary up/down, train a regression model on return_EOD directly, then threshold at trading time. Or use a two-stage approach: classifier for direction + regressor for magnitude. This captures more information from the target and can improve calibration.

I had high hopes for this, but actually it was roughly comparable or slightly worse than the best binary classifier.

Still to try...

More historical data — Going back to 2023. More data always helps.

Better LLM features — use a more capable model for extraction. The feature extraction uses gpt-5-mini. Upgrading to gpt-5 (or even gpt-5.4) for a subset of key features might extract more nuanced signals.

More LLM features — Short interest changes pre-earnings, social media sentiment, news volume. All require new data sources.

Sector-specific models. The sector AUCs vary significantly (Utilities 0.72, Basic Materials 0.55). Training dedicated models per sector group (e.g., Tech+Comms, Healthcare, Financials, Everything Else) could capture sector-specific patterns that a single model averages out.

Ensemble of diverse model types - XGBoost and LightGBM are similar (both Gradient Boosted Decision Trees). Adding a genuinely different model — a neural network, or a linear model with different feature engineering — into an ensemble could capture complementary signal.

I’ll also keep an eye on those 28 day models , I could pursue a strategy using options that would be capital efficient and perhaps better alpha.

OK - slightly boring, but don’t worry, I have some more fun stuff to come.

I’m hoping these marginal improvements give me a model that generates alpha whilst I work on everything else (and avoid going broke. )

This means that those features revealed or suggested what actually happened to that quarter, making it easy for the model to guess what happened… It was as if the exam answers were stapled to the question paper. Or at least, clues to the answer.

The calculate_earnings_move_stats method included the current quarter’s outcome when computing these stats, meaning the model was indirectly seeing the target variable. We fixed that by ensuring this feature did not include the current call.

After fixing this all models were retrained.

1. The intraday models (2h, 3h, 4h) were the most inflated :( — they dropped 0.10-0.12. This makes sense: earnings_win_rate (a stock’s historical tendency to go up after earnings) was directly leaking the current quarter’s outcome, and shorter horizons correlated more tightly with it.

2. EOD models barely changed (-0.016 to -0.018). The EOD target has more noise from full-day trading, so the leaked feature was less predictive of it.

3. 7d was negligible (-0.005). Over a week, the leaked feature adds almost nothing.

4. 28d actually improved (+0.010). Cleaner data helped the model generalize better.

5. All models now converge to the 0.62-0.64 range. The “4h is dramatically better than EOD” finding was fake. In reality, the intraday and EOD models perform similarly.

6. earnings_win_rate dropped from #1 feature to #9 in the 4h model (importance 0.037 → 0.008). It’s still useful, just not dominant.

The resulting correlations still look promising though.

An example analyst from my old approach…

…versus one of the latest XGboost EOD models…

Time to trade

Once the leak was fixed, I set the most promising model off trading. In the meantime I’m off on holiday.

Can AI do that?

Part 3 tomorrow : a trading disaster and marginal gains.

Also - I am not happy with these videos!!! They will be improving.

Anyway….

My old approach was to ask the LLM for a buy confidence based on the earnings transcript, with different prompts highlighting different signals. The problem was that LLMs are not great at returning this type of numeric rating. And the lack of correlation I was seeing proved that…

“The new system” is quantitative signal generation or alpha signal modelling — using structured features derived from earnings calls, technicals, and fundamentals to predict short-term price direction.

What LLMs are very good at is extracting information from text. Reading 25,000 transcripts and creating around 200 features from each took a couple of days and cost less than $100, with gpt-5-mini anyway.

Example features are

• guidance_raised: Did management explicitly RAISE revenue or earnings guidance for next quarter or full year?

• revenue_declining_segment: Is revenue declining YoY in one or more major business segments? Must be explicitly stated as declining versus prior year.

• beat_own_expectations: Did management state results EXCEEDED their OWN internal expectations or prior guidance? Look for “above the high end of our range”, “exceeded our own expectations”, “ahead of plan”. Beating consensus alone does NOT count.

And there are also macro market features for that day:

• macro_risk_off: Is pre-market sentiment risk-off? True if overnight futures are notably down, safe-haven assets (bonds, gold, yen) are rallying, or there is a clear flight-to-safety tone in pre-market coverage. Normal fluctuations are false.

• fed_or_rate_event: Is there a Federal Reserve event today — FOMC decision, Fed minutes release, Fed chair speech, or any major central bank rate decision (ECB, BOJ, BOE) that could move US markets? Scheduled events only.

• geopolitical_escalation: Has there been a significant geopolitical escalation overnight — military conflict, new sanctions, terrorist attack, political crisis, or diplomatic breakdown that is leading pre-market news coverage and could materially impact markets?

• vix_elevated: Is the VIX (or expected volatility based on pre-market conditions) notably elevated — above 20, or spiking significantly from recent levels? Normal VIX in the 12-18 range is false.

Finally, with have some traditional quant features and other metrics :

• delta_sma_200_50 - simple moving average of past 200 days vs more recent 50

• hist_avg_return_eod - average EOD return

• qa_word_count - how many words in Q&A portion of earnings call

So every quarter now has the following…

Next I use traditional supervised learning machine learning, supervised in this context meaning I gave the model what the result was (did it end up or down) and it tries to learn how to predict it.

To start with I’m using gradient-boosted decision trees (GBDT) for binary classification ( i.e. there are only two possible predictions, stock will go up or go down).

Two key metrics

AUC (Area Under the ROC Curve) measures how well the model separates stocks that go up from stocks that go down, across all possible confidence thresholds. An AUC of 0.50 means the model is no better than flipping a coin — it can’t distinguish winners from losers. An AUC of 1.0 means perfect separation. In financial markets, where randomness is extreme and most signals are weak, anything consistently above 0.60 is meaningful edge.

Another metric is Win rate i.e. of all the trades the model actually recommends (above a chosen confidence threshold), what percentage ended up profitable? For example, at a 0.70 confidence threshold if we have a 68.6% win rate that means 7 out of every 10 trades close up. Higher thresholds give better win rates but fewer trades: at 0.70 we get 433 trades with 68.6% wins, while at 0.75 we get 293 trades with 72.7% wins. The art is finding the sweet spot where win rate, number of trades, and average return per trade combine to maximise total profit after transaction costs.

Baselines

Initial models trained on ~21k rows with time-based splits. EOD achieved AUC EOD 0.649, 7d 0.612, 28d 0.615. Solid but modest.

Micro-cap stocks were the most difficult to predict so in the end I excluded them for now.

Next decided we widened hyperparameter search and switched to LightGBM, a more advanced GBDT that was developed by Microsoft in 2017.

We also added 57 new features (volatility metrics, company earnings stats, financial ratios, insider/analyst data).

EOD AUCs stayed around 0.643-0.645. We then trained intraday models (2h, 3h, 4h) which appeared to dramatically outperform — the 4h model showed AUC 0.750…. but there was a big issue…. for next time.

Part 2 tomorrow.

To my new readers and folks I’ve not spoken to in a while, here is a general introduction to my latest misadventure.

• Why I set out on this Sisyphean journey

• How I use machine learning to try predict stock movements

• Why I think it will work

To be continued…

We are not at war.

An alcoholic Fox News host, who now runs the Orwellian Department of War, which itself answers to the lifetime chair of the Board of Peace, who has bombed seven countries so far this year, has threatened to cut out or even nationalize the world’s most advanced AI company… How much we really do live in free market capitalism?…

Money Monster’s does keep an eye on the economy but actually most active analysts don’t use prompts that factor it in. Here was MM’s summary today…

…the U.S. economy is facing significant challenges, primarily due to escalating geopolitical tensions in the Middle East, particularly the ongoing conflict with Iran. This situation has led to a sharp increase in oil prices, with Brent crude reaching $85.41 per barrel, the highest level since 2024. The surge in oil prices has heightened concerns about inflationary pressures and the potential for prolonged economic strain…

It goes on to summarise the effects on the energy, tech , consumer goods , banks and airline sectors.

It’s likely that the classic F150 driving MAGA voter is going to have an unpleasant surprise next tie they visit the pumps.

But how did MM do? It was a mixed bag.

A rocky week

Today

Yesterday

A note on the numbers below - the first number is the change from the market open, which isn’t always a great indicator given that calls might finish after market open. The second number is the actual position return that day. What it does show yesterday is that if these positions had been left to run - they would have had a good return.

So with dynamic brackets for stop loss and take profit, which take into account a company’s typical volatility on earnings day, the ALTO position would have stayed open.

Wednesday

A dystopian week

… with a few dystopian stock picks

May you live in interesting times.
A

The story of today, which was a Green Day with all stocks up.

My main complaints are that not much of the buying power (roughly 3x actual deposits) was used, which is inefficient. And I worry that the conservative sizing logic is holding back some of the best performers. All to figure out.

But… today was the first morning I can remember with no warnings or emails from the machine, where I sat reading in the garden until 10am before visiting the office to see what Money Monster was up to.

Some thoughts whilst I’m feeling philosophical.

An idea sounds easy

A simple phrase doesn’t belie the million obstacles on the way.

“Do an AI to trade earnings calls” doesn’t exactly sound simple, but doesn’t reveal so everything that has to go right. This is true of any business, project, or endeavor.

Each morning this system tunes into all 25+ earnings calls, detects their completion, tracks financial metrics, summarizes the transcripts, passes them on to the analysts, queues up the trades, sends them for execution and reconciliation, and tracks throughout the day.

There have been so many frustrating mornings with things not working, missed trades or botched analysts, simple bugs, or even just the time it takes to build something, to make something real from your imagination.

A common response I got to this project is “aren’t people doing that already?”

Well yes kind of, but doesn’t mean one more can’t join them, especially with a few friends (Opus 4.6, GPT 5.2). It’s helpful when your friends have a pretty decent grasp of all human knowledge and the reasoning capabilities of a supercomputer from a an old sci-fi movie.

But I also think that only through the friction of being fully engaged with something can you begin to understand the opportunities. That’s why so many successful businesses are started by someone that was previously in that world and could see the opportunity.

But my bigger point is: things are achieved by the people that push ahead regardless of the chances. Founders, not that they are the only people that do this, but they are the ones I know, always have this slightly deranged zeal that propels them forward. Who knows if they will make it? We don’t - but we know that they certainly won’t if they don’t try.

Timing

Treading water in the right place is so key to a project like this. No matter how many hustle-bros tell you it’s all about grinding, the reality is that it’s about a lot more than that, and one of the most capricious ingredients is luck.

The harder I work, the luckier I get, as the old adage goes. Well, that’s kind of true. If you are going in vaguely the right direction this certainly speeds things up. But how key is luck? I often think about documentaries, the best ones out there are the ones where something crazy happens that the producers could not have possibly predicted.

Not as big as that, but this month Opus 4.6 was released by Anthropic. I was instantly building faster and smarter as soon as it be available than before.

Agentic ‘end to end’ AI takes over the hype cycle, but for most of us on the tools AI functions more as a cognitive exoskeleton, helping us ‘lift’ more and get more done. It is remarkable how much more productive I am in an hour than I was three years ago, and that’s not just because I need to make some money.

I still have to eat, exercise, sleep, get drunk, all important human endeavors, they are becoming the limiting factor which will bring us to…

Agentic AI

When you hear about people leaving their machines to code all night whilst they sleep, they are talking about agentic AI. You give it a goal and it does the rest, chooses the tools, remembers the context, adjusts based on results.

What would a true agentic AI software build given this prompt: “Build a service to trade earnings transcripts intraday for profit. “ ? It would be interesting to watch, it would probably come up with ideas that I haven’t yet. Thought I do think human taste and imagination have some value yet.

Better, faster, cheaper

One of the reasons I knew that crypto would not live up to its promises is that good technology should make things better, faster, and cheaper. Crypto doesn’t do that, except in very specific use cases like crime.

However AI is useful. These frontier models are surpassing our expectations every few months. If you used AI in the past and were not impressed, you need to try again now. Everything is better from reasoning, speed, multitasking, context size, it does more for you than before. And the previous state-of-the-art models are getting cheaper and cheaper.

Paying for Cursor or OpenAI is the fastest I’ve handed over my credit card. Cursor hit 100 million ARR faster than anyone in recent memory (credit: Aakash Gupta) because the value is obvious.

Learning

I did study AI in 2008 aka ancient history, but I tried to do the Harvard CS course more recently to get up to speed again. Of course it cannot possibly keep up with the state of the art, although it does teach the fundamentals of modern neural networks.

However, having a project like Money Monster has far surpassed (and made more fun) what I would’ve learned on a course. Applying it to my own projects and understanding how it helps with each objective has the emotional resonance to keep me learning. And to stay inspired.

Software engineering and AI

I generally avoid the dystopian topic of the future, especially as some parts seem to have arrived early. But it’s not lost on me that I used to run a business with 100 employees. Now I am buying computer power. And how many software developers does it need to change a light bulb now? I was struck in this video from a documentary about DeepMind and Demis Hassabis how antiquated this description of a software developer seems. What is the right description today ?

Signing off for the weekend , I’ll leave you with a final note from my friend Atlas on the topic of AI job displacement just today.

Money Monster sums it up ‘We’re up 1.77% today (+$3,108) and giving SPY a wedgie while it’s down 0.57% — cashing out clean with no open positions.’ That’s good - but the past few days have been a bit up and down, with some technical issues and but more importantly, lots of picks that moved in both directions, resulting in not much movement overall, and lots of risk.

The good thing is that the sizing from last week came into effect which reduced some of the bad losses, but also limited some of the wins. Overall the fund didn’t move much until today.

I’m leaving the fund side for a moment (sizing and fund allocation) and I’m going to focus on the analysts. I’d been hoping to put out a video on how I make them, but didn’t create any new good ones and it felt a bit flat. (I am doing my best to try and make ‘man talks to screen’ interesting.)

Thankfully I’d arranged to chat that day with my friend Sam Bell who is a Machine Learning PhD and until recently an AI Research Scientist at Meta. So if he can’t help…

Sam’s idea was to focus move towards learned inference. The current approach is a rule-based system (or expert system). The decision rules are hand-crafted by the author of the prompt, which last year was me, but this year is normally Anthropic’s Opus 4.6, a frontier Claude model released earlier this month, which has certainly improved the phrasing and creativity of prompts.

But the criteria the prompts look for doesn’t change or improve unless I decide to change them. Any learning they do comes by comparing quarters within a given prompt and is ephemeral. This doesn’t really mean Money Monster is ‘learning’ over time.

The LLMs I use (normally GPT 5 series) generate a score, but they can also return features e.g. sector, # of employees, regulatory approvals, analyst reactions, the results look like this:

  “revenue_growth”: “ACCELERATING”,
  “guidance”: “NOT_GIVEN”,
  “cash_position”: “STRONG”,

These become manually tuned heuristics e.g. ‘if revenue_growth is positive , + 15 points’

The new approach is supervised machine learning. “Supervised” because I give the model labelled examples which consist of :

• the features but preprocessed for comparison across all quarters

• known outcome - did the stock go up or down

The decision weights are learned from data rather than specified by a human.

You can think of the scale normal programming to modern artificial intelligence as:

Money Monster is moving from row 2 to row 3.

Row 4 (deep learning, neural networks, real “AI”) would be fine-tuning a language model directly on transcripts to predict stock direction -- no feature extraction step at all. That’s the most powerful but needs the most data and compute. In the meantime we can :

• establish a baseline to improve upon

• build the infrastructure we’ll need either way e.g. features table

• work out if the features have any signal at all

• focus on which are the best features before expensive GPU time for fine tuning

Let’s hope Money Monster can get smarter faster.

Lots of lessons, implemented immediately

The positions yesterday then:

Let’s focus on the negatives (Why break the habit of a lifetime? ) MATV was the bombing position I start the video with

Two problems here:

• The buy confidence threshold for this analyst was too low.

• Position size was equal to the others, even though this one is more volatile.

• Related, buying power bugs meant that other (good) positions didn’t open.

Explanation from my AI friend:

Today the system opened three large positions (TS, MATV, TALK) totaling ~$186K in rapid succession at market open, which drove the Reg-T buying power down to $0 while the daytrading buying power still showed ~$51K. Because the sizing code was reading the wrong buying power field (daytrading instead of Reg-T), it sized two more positions (TFPM and FTI) at $51K each — but Alpaca rejected every order since Reg-T was actually zero. Worse, the insufficient buying power check was advisory-only (just a warning log, didn’t block the order), and there was no retry limit, so the system hammered Alpaca with doomed buy orders every 10 seconds for over 4 hours, generating 1,902 error log entries. FTI eventually squeezed through around 9:51, and TFPM finally opened at 12:11 after MATV’s stop-loss freed up buying power — but by then TFPM’s trade value had shrunk from $51K to $27.73 (1 share). The fixes use the correct Reg-T buying power for sizing, block orders when buying power is insufficient instead of just warning, and give up on positions that can’t open after 30 minutes.

All deployed.

Sizing & Volatility

My friend Feederico works in Real Finance™ and warned me (from day one) that volatility should always be factored into position size.

Like so many times in life, I had to learn my preferred way, the hard way.

Currently, volatile and calm stocks get the same dollar exposure, meaning a single bad earnings reaction on a volatile name could cause an outsized loss relative to the rest of the portfolio i.e. what happened yesterday.

The system added volatility metrics (ATR, historical volatility, and average earnings move) then simulated yesterday’s positions. The MATV position was cut from from $62k to $16k, reducing its loss from -$6,035 to -$1,534 and improving the day’s total P&L from -$3,259 to -$455. Still a loss but a mild one.

Yesterday with volatility-adjusted sizing

Magic.

Let me know if any more feedback or ideas, whether on MM or this here content (also is my mic loud enough??)

There were no buys today but that’s ok, I’m happier with these upgrades done.
Let’s hope for more luck next week.

Tuesday & Wednesday….

The trades themselves are below. Note that the first number is the delta from market open, whereas the second bolder one is the difference between the buy decision and the actual position close i.e. what the actual return was.

Unexpected benefits whilst working on this

• I mainly speak to my computer now using Monologue, and it makes Apple’s dictation seem Stone Age. However, the consequences dictating is that I find myself making more effort to enunciate properly, which was a goal of mine a few years ago. Also, speaking is one of the best ways of getting your thoughts in order. It turns out the listener doesn’t have to be human.

• My thinking is now shifting to how quickly can I give this task to an AI to work out? Some people might worry that is going to disable them in some way. I’m willing to trade that for the fact that I am at least 3x more productive and will always need our internet connection to work. Fine.

• I am becoming the blocker. The multitasking, given how many different agents and tasks can be running at any one time, is becoming one of the most difficult parts of this job. I find myself exhilarated and exhausted at the end of the day, being the conductor of these cognitive wizards. That can be dizzying, but also means the potential to create in this industry is wild.

• It’s making me think a lot about the nature of cognitive work, which I’ll write about soon. Just how in software we’ve always had 10 ex-engineers and then mid ones, we’re probably going to have the same here with operators i.e. those that can grasp the tools will be able to create magic. Those stuck pre-2020 are going to be too slow. .

• Just like how when there’s a small internet lag, I find it frustrating, despite memories of waiting five days for a song to download, I have the same now with ChatGPT and other LLMs. The small wait as it thinks, I’m already wishing for when they have GPUs powerful enough to make that instant. It was only three years ago this all seemed magical. And now I’m complaining about waiting five or ten seconds for an elegant and incisive answer on basically any topic ever.

• At 41, I used to think that maybe my own thinking was getting slower, and certainly I hadn’t been on the tools for a long time whilst I was running my business. Amazingly, now that doesn’t matter. I missed out on a whole decade of JavaScript taking off a front-end web design, Which I never fully understood being a good old-fashioned PHP/Java/Ruby HTML developer. Now , blissfully, I’ll never need to know.

MoneyMonster has its own IG @moneymonsterfund

Follow along , it has a bit of character. this update from my development environment (which has stale data) made me laugh.

Where I’m at with this

The premise here sounds simple “ Trade earnings calls with AI. “, but the reality with all businesses is that it’s the million small details, mistakes, decisions and challenges along the way, which means perseverance is always key. I’m still not there. But I have to admit today felt like a bit of a milestone. I had the idea of MoneyMonster having its own IG over a year ago, and I finally got to set it up today. The outputs you see are built by Opus 4.6 (architecture) with GPT-5 for MM thoughts/voice/reasoning.

There are so many optimizations to make with this system, including the many areas of slippage such as transcripts being late, processing taking too long, and actual positions not opening in time.

That’s okay, the key thing at this point was getting a proof of concept.

Next is optimization.

Then research different areas ( Trading the news, foreign equities markets, going deep on one sector , Ideas welcome here!! Hit reply )

If the core premise is correct i.e. that LLMs can ‘reason’ as good or better than a Human analyst, which is what a multi-trillion dollar industry is based on, then this should work eventually.

And even if that’s not the case today, what stops it being the case in six months?

Your human in the loop, signing off.
A

This week the approved analysts have been making some good decisions. There were no serious failures or bugs this week, the system listened into around 50 calls at a time each morning and executed the orders for stocks it liked.

Yesterday was a bad day

Yesterday I accidentally let a rogue analyst loose.

What conclusions?

• More patience, avoid tilt, which is “a poker term for a state of mental or emotional confusion or frustration in which a player adopts a suboptimal strategy, usually resulting in the player becoming overly aggressive” If I just let the system I set up last week do its job, we’d be up a lot this week.

• Better change control for new analysts e.g. a staging environment where they can run in parallel.

• Run backtests over longer periods… $1000 on OpenAI credits would’ve saved me much more here.

• The bracket orders (take profit and stop loss) worked well but the take profit can be increased from +15% to +20% … when we have winners let them ride.

How AI has changed my days

I am pretty much only talking to my computer now using Monologue, which is instant and about 10 times better than Apple’s default transcription.

Whilst I’ve been writing this email, I have an AI looking for my new health insurance via OpenAI’s Atlas browser. Another one working on a complete refactoring of another system in Cursor. And one helping me with future writing.

In coding alone, I have been relegated to a mere project manager (or visionary??) with an expensive team of AI engineers working concurrently. With Money Monster there are effectively 10 analysts listening to and acting on each earnings call.

The abstraction is getting higher and higher, I’ve gone from asking for very specific changes (senior engineer)… to providing a high level spec (product manager)…. to just stating the outcome I want (CEO). With no employees.

Lots of good thoughts this and more in this article :

• “The future is being shaped by a remarkably small number of people: a few hundred researchers at a handful of companies... OpenAI, Anthropic, Google DeepMind, and a few others. A single training run, managed by a small team over a few months, can produce an AI system that shifts the entire trajectory of the technology. “

• People in tech are already realising “I am no longer needed for the actual technical work of my job.”, this was different to six months ago

• If you didn’t think AI worked when you tried it, or that it made mistakes, you need to retry the current models and not use the default free ones - even in the last six months their capabilities have transformed.

• The major AI companies are now using their own state of the art models to recursively build the next versions of these models, so expect progress to accelerate.

• My focus here is around productivity rather than the implications for society, the economy, national security and so on, which Jasmine Sun recently went deep on in her piece about AI populism vs big tech. Somewhere in a building like this in Ohio there is a rack doing all my work.

“Imagine it’s 2027. A new country appears overnight. 50 million citizens, every one smarter than any Nobel Prize winner who has ever lived. They think 10 to 100 times faster than any human. They never sleep. They can use the internet, control robots, direct experiments, and operate anything with a digital interface. What would a national security advisor say?”
—Mark Amodei, U.S. representative for Nevada’s 2nd congressional district

Final note - the siren in the video was for a fatal accident outside my house, a woman on her phone was run over by a careless driver. He wasn’t speeding - he was also on his phone. Stay safe out there.

A

The goal: EOD price prediction (2 mins)

Pre-market setup, model/analyst checks, position approval, and resulting intraday performance, this is the dream.

How MM & I generate new ideas (6 mins)

A deep dive into using LLMs to detect earnings beats from transcript and financial signals. Using Anthropic Opus to help engineer the system, including the prompts themselves (self learning?), to develop and backtest new ideas in under 10 minutes. Definitely a better idea than experimenting live in production….

(My tardy emails. Content production speed is improving, aiming to improve the quality and speed of these videos in the next couple of weeks. And hopefully have more good stock picks each day. )

My broker had a lot of pricing issues this morning and it was amazing watching money monster handled it all, not throw any errors and instead make a couple of profitable trades. More in the video.

You can read the rational for the buy decision and the aftermarket retrospective from money monster below.

Not a bad day one - a few lessons and refinements. Checking if the ‘pop’ has already happened for transcripts that MM is late to process , but for now just not trading too far past market open. I have a lot to say - but I’ll save it for another day. Stay tuned.

Here lays the premise of the dream.

I can hunt in the morning, fish in the afternoon, rear cattle in the evening, criticise after dinner, just as I have a mind, without ever becoming hunter, fisherman, herdsman or critic.
- Karl Marx

Today’s Buys

These trades are made at market open 9:30 Eastern. Video is from 9:35am. So I cheated a little - but couldn’t resist watching 8x8 go up 16%, let’s hope it stays that way.

1. EGHT — 8x8

   a company that sells yet another cloud calling, conferencing, and contact-center subscription nobody needs
   Micro-cap Technology

   Q3 was a clear beat: total revenue ($185M) and service revenue ($179.7M) both topped the high end of guidance by ~$3M, operating margin was 11.7% vs. 9–10% guided, EPS beat the high end by $0.03, and operating cash flow exceeded guidance.

2. WFRD - Weatherford International,
   a company that extracts fossil fuel while promising sustainability and failing.
   Mid-cap energy.

   Q4 showed solid execution (rev +5% q/q, EBITDA margin 22.6% +74 bps, $222M FCF), low net leverage (0.42x), and stronger shareholder returns (10% dividend hike, buybacks), which are typically supportive. However, Q1 guidance reflects seasonality and 1H softness, FY26 margin guide is slightly below Q4, and Mexico collections remain a watch item. Net effect: modestly better-than-even odds of an up day.

3. JCI - Johnson Controls International
   a company that outfits buildings with costly climate-control systems that rarely behave.
   Large-cap industrials.
   Q1 FY26 was a clear beat on profitability and slightly ahead on revenue. Organic revenue grew 6% and management said the quarter came in ahead of expectations, while adjusted EBIT margin expanded 190 bps to 12.4% and EPS of $0.89 rose ~40% and exceeded their guidance. Orders surged ~40% and backlog rose 20% to a record $18B, supporting visibility. T

4. That’s it so far but there are still 25 calls in progress.

   ---

   What is Money Monster?

   MoneyMonster is an automated trading system. There is no human decision making. The input is language (earnings transcripts back to 2016) and price movements. The system then decides whether to buy or not.

   The rules

   • Universe: U.S. equities (NYSE / NASDAQ / AMEX), filtered for liquidity.

   • Holding window: intraday only. Entries after the open, exits 15 minutes before the close.

   • Positions per day: typically 0–5, often none, occasionally more.

   • Sizing: equal-weight “units” with a max cap per name.

   • No discretion: I don’t override anything. There are stop loss (-10%) and take profit (+15%) orders outside of the AI.

   What to expect

   • Some days: the day’s trades (tickers + direction + position unit + short rationale)

   • After-close recap: results + post-mortem

   • Weekly: performance context + model changes + what I learned

Thanks for being here on Day 1. I’ll be back after the close with what actually happened.

Subscribe now